Tool Permission Policies
The problem, from first principles
An agent that can run shell commands, send emails and delete files is only as safe as the rules that stand between a model's output and the real world. A permission policy is the simplest effective guardrail: a list of patterns saying which tool calls are allowed outright, which are forbidden and which need a human to approve first. Two design rules make it robust. Deny beats allow: an explicit prohibition cannot be overridden by a broad allowance. And the default is deny: anything the policy does not mention is blocked, so forgetting a rule fails safe instead of open.
From theory to code
Implement check_permission.
Constraints
callis a string such as'bash:rm -rf /tmp/x'or'email:send'.policyis a dict with keysdeny,askandallow, each a list of glob patterns (usefnmatch.fnmatchcase).- Return
'deny'if anydenypattern matches. Else return'ask'if anyaskpattern matches. Else return'allow'if anyallowpattern matches. Otherwise return'deny'. - Missing keys in
policymean an empty list.
Hints
Hint 1
Test the three lists in priority order and return at the first hit.
Hint 2
fnmatchcase('bash:rm -rf /', 'bash:rm *') is True, since * matches any characters including spaces and slashes.
Keywords
No results yet
Run your code to check it against the examples, or submit to run every test.